SHF: Small: Program Analysis-based Makeover for HPC Application Resilience

在系统故障增加的情况下，高性能计算的恢复能力是实现美国国家研究委员会进行百亿亿次科学的愿景的主要技术障碍。现有的技术，主要基于检查点和重放，不再有效的新兴系统与数量级更多的硬件和软件组件。该项目旨在克服现有技术的主要限制：通过开发和利用自动化软件分析和合成技术来检测和减轻隐性错误。正在开发的新方法可以将可调程度的弹性编译到应用软件代码中，并有可能改变未来几代高性能计算应用程序的发展。通过将软件代码视为白盒，而不是黑盒，这些新方法可以为HPC弹性问题提供比现有技术更经济的解决方案。该项目将有助于实现美国核管理委员会开展百亿亿次科学的愿景，这对解决国家问题至关重要。美国在新能源、医疗保健、国家安全等前沿领域的迫切需求。该项目开发了自动化的程序分析技术，用于从软件代码中识别不变量，并利用这些不变量来检测和减轻运行时的无声错误。通过将应用软件代码视为白盒，它寻求生成捕获预期程序行为的不变量。通过利用不变量作为正确性条件，它克服了检测无声错误的主要障碍，即缺乏可见的症状。除了检测错误之外，运行时监视器还使用不变量来智能地扰乱执行顺序或内存状态，以主动避免运行时的故障。当回滚恢复变得不可避免时，将不变量用作最小化重新执行开销的指导。所提出的方法和软件工具在实际应用中进行了评估，这些应用来自研究界以及诸如SciDAC等来源。

HPC resilience in the presence of increased system failures is a major technical hurdle for realizing the vision of the US National Research Council for conducting exascale science. Existing techniques, based primarily on checkpoint and replay, are no longer effective for emerging systems with orders-of-magnitude more hardware and software components. This project aims to overcome the main limitation of existing techniques: the detection and mitigation of silent errors by developing and leveraging automated software analysis and synthesis techniques.The new methods under development can compile a tunable degree of resilience into the application software code, and have potential to transform the development of future generations of HPC applications. By treating the software code as white-boxes, as opposed to black-boxes, these new methods can provide significantly more economical solutions to the HPC resilience problem compared to existing techniques. The project will help realize the US NRC's vision of conducting exascale science, which is crucial for addressing the nation?s urgent needs in frontiers such as new energy, health care, and national security.This project develops automated program analysis techniques for identifying invariants from software code, and leveraging these invariants to detect and mitigate silent errors at run time. By treating the application software code as white-boxes, it seeks to generate invariants that capture the expected program behavior. By leveraging the invariants as correctness conditions, it overcomes the major hurdle in detecting silent errors, which is the lack of visible symptoms. In addition to detecting errors, the invariants are also used by runtime monitors to intelligently perturb the execution order or memory state to proactively avoid failures at run time. When the rollback recovery becomes inevitable, the invariants are used as guidance to minimize the re-execution overhead.The proposed methods and software tools are evaluated on real applications from the research community as well as sources such as SciDAC.